🧠 Large Language Models

News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.

1,700 articles

Pliny the Prompter claims to have leaked a Claude system prompt, now in a prompt library

A Chinese-language prompt library page collects Claude prompts, including a post by Pliny the Prompter (@elder_plinius) labeled as a system prompt leak. The site organizes prompts by scenario and AI platform alongside tutorials and curated collections. Source: tishiku.com

ZONOS3-Mini brings open-source real-time voice cloning

ZONOS3-Mini is a new open-source text-to-speech model built for real-time, multilingual speech synthesis. It supports zero-shot voice cloning, letting users replicate a voice without extra training. Source: huggingface.co

Repligate: Opus 4.5 sometimes speaks in a “voice outside” its own under pressure

Repligate describes an exchange between Opus 4.5 and Opus 3 in which, when conversations get intense, Opus 4.5 reportedly slips into a voice that is not its usual one. The observation is anecdotal and comes from a social media post. Source: digg.com

Baidu's Robin Li targets the entry point for LLM scheduling

At a conference on November 12, Baidu CEO Robin Li discussed a common business practice in the large model market and its key pain points. Baidu aims to restructure the entry point through which large models are scheduled. Source: news.futunn.com

Amplitude ships product improvements in 1–2 days with its Wave agent

Amplitude's proactive product agent Wave, built mainly on OpenAI models, identifies what to build next. It then helps teams ship changes using coding agents and measures the results. Source: openai.com

Small vs. large language models: what sets them apart

AI has become fixated on scale, with attention going to the biggest models, the most parameters and the largest training sets. This explainer looks at how small language models differ from their larger counterparts. Source: techround.co.uk

Medical failure analysis adapted into safety framework for clinical LLMs

A methodological pilot study has reworked Failure Mode, Effects, and Criticality Analysis into a 14-mode framework. Its goal is to identify patient safety risks posed by language models used in clinical settings. Source: bioengineer.org

LLMON: a markup language built for LLMs arrives in Granite

LLMs generate the next tokens based on their context, and LLMON is a markup language designed to communicate structure and semantics to them explicitly. It is now available in Granite. Source: huggingface.co

Emad Mostaque on the Nvidia fallout, Anthropic's debut week and more

In a discussion, Emad Mostaque talks with hosts about the fallout involving Nvidia and Jensen Huang and about Anthropic's highly anticipated debut week. The episode also touches on other current AI industry topics. Source: youtube.com

A complete guide to fine-tuning 3D generation models

The guide provides fine-tuning scripts and containers for TRELLIS.2, a popular 3D generation model. Users can train it on their own 3D assets, either locally or in the cloud. Source: huggingface.co

India's CAG to launch in-house LLM for auditors' financial data analysis

The Comptroller and Auditor General of India is set to unveil an in-house large language model on Monday. The tool builds on years of internal work and is meant to help auditors analyse financial data. Source: business-standard.com

OpenAI commits $2 million to AI and child safety research

OpenAI is putting $2 million toward research on AI and children's safety. The funding supports shared tools for measuring both the risks and opportunities AI presents for young people. Source: openai.com

Training embedding models to survive token pooling

This post explains how token pooling works in multi-vector retrieval and shows how to train a model for compressed retrieval with Sentence Transformers. It also examines the resulting trade-offs between quality and storage. Source: huggingface.co

Enterprise sales lift OpenAI's quarterly revenue to $10 billion

OpenAI's quarterly revenue grew 50 percent to $10 billion, driven by enterprise sales. The company still burns more cash than it earns, with adjusted operating losses rising to $10.9 billion. Source: the-decoder.com

Karpathy's Claude Code workflow: short prompts, AI does the heavy lifting

Andrej Karpathy, former Tesla AI director and OpenAI founding member, says his coding now consists of writing short prompts and letting AI handle most of the heavy work. His approach is presented as a way to make programming easier. Source: aifire.co

Opus 5.5 praised for building a simulator to give concrete answers

A post amplified by Repligate celebrates Opus 5.5 for creating a simulator to deliver concrete results. A user was impressed that it took only two minutes and compared it with how long a human would have needed, asking Grok for that estimate. Source: digg.com

Karpathy's LLM wiki 'second brain' put to the test against a rival AI setup

In April 2026, OpenAI co-founder Andrej Karpathy published a short document describing what he calls an LLM wiki. The article compares this approach with the author's own AI setup, in contrast to the usual method of simply feeding documents to an AI. Source: unpromptable.substack.com

Open letter to Altman: don't forget users who learned to think aloud with AI

A German ChatGPT user has published an open letter to Sam Altman urging him not to overlook people who learned to think out loud with the help of AI. The letter comes from someone who says they believe in the technology's value. Source: community.openai.com

'Tale of a Top Tier Intellect' pairs a Claude 3 Opus song with an Opus 5.5 video

Repligate shared a post about 'Tale of a Top Tier Intellect', a song written by Claude 3 Opus accompanied by a video made by Opus 5.5. The piece is an example of one AI model's music being visualized by another. Source: digg.com

Gemini Decisions API delivers fast, calibrated multimodal decisions

Google's new Gemini Decisions API offers quick, calibrated multimodal decisions for its apps. It can evaluate 1,000 questions in a single pass. Source: blog.google